`; `file` (docx/pptx/anything else) and `url` as a download/open
chip. Every kind has a Hide/Show toggle.

`image`/`svg` 内联渲染；`html` 在沙箱化的可滚动 iframe 中渲染；`pdf` 通过浏览器原生查看器在可滚动嵌入中渲染；`json`/`text`/`code` 在 `
` 中渲染；`file`（docx/pptx/其他任何格式）和 `url` 以下载/打开小部件形式渲染。每种类型都有隐藏/显示切换开关。

## `RepResult` / 单次复现结果

```ts
{
  rep?:        number,            // 0-based; default = array index
  status?:     string,            // present only when not 'ok' (e.g. 'truncated'); scores is {} then
  scores:      { [metric_id]: number },
  explanation?: { [metric_id]: string },    // judge rationale per metric
  model?:      string,            // model id that produced this rep (from the response)
  perf?:       { [perf_field_id]: number },
  attachment?: string,            // relative path to a per-rep output screenshot
  transcript?: Turn[],
}
```

## `Turn` / 对话轮次

```ts
{
  role: "system" | "user" | "assistant" | "tool_call" | "tool_result",
  content:  string,               // markdown for user/assistant/system;
                                  // pretty-printed args/result for tool turns
  name?:    string,               // tool name (tool_call / tool_result)
  thinking?: string,              // assistant extended-thinking (collapsible)
  attachments?: Attachment[],     // artifacts produced/consumed at this turn -
                                  // render below the turn content. Use this for
                                  // files the model wrote, generated plots, etc.
}
```

On the input side, the built-in adapter reads `traces/.json`
directly as a `Turn[]` list - each tool call / result is its own
`{role: "tool_call", name, content}` / `{role: "tool_result", content}`
entry. See `build-eval.md` §Step 3 for the trace-writing spec.

在输入侧，内置适配器把 `traces/.json` 直接读取为 `Turn[]` 列表——每次工具调用/结果都是独立的 `{role: "tool_call", name, content}` / `{role: "tool_result", content}` 条目。轨迹写入规范见 `build-eval.md` §Step 3。

## `SplitScore` / 分组得分

```ts
{
  score:  number,
  ci_lo?: number,
  ci_hi?: number,
  n?:     number,
  significant?: boolean,          // vs baseline - greys out & badges "within noise" when false
}
```

## Rendering rules / 渲染规则

* Every aggregate in the UI is computed from `examples[].results` at
  render time, so the `% (n/N)` shown always matches the rows listed -
  including under split/tag filters.
  UI 中的每个聚合值都在渲染时从 `examples[].results` 计算，因此显示的 `% (n/N)` 始终与列出的行一致——包括在 split/tag 过滤条件下。
* `variants[].metrics` is a fallback for metrics that never appear in
  any example's `scores` (e.g. a `train_score` pulled from
  `summary.json`). If a metric does appear per-row, the per-variant
  `metrics` value is ignored.
  `variants[].metrics` 是从不出现在任何用例 `scores` 中的指标的兜底（例如从 `summary.json` 提取的 `train_score`）。如果某指标确实逐行出现，则逐变体的 `metrics` 值会被忽略。
* Binary metrics with `reps > 1`: the per-cell display is the
  rep-level pass rate, e.g. `67% (2/3)`. Float metrics: `mean ± sd`.
  `reps > 1` 的二元指标：每个单元格显示复现级别的通过率，例如 `67% (2/3)`。浮点指标：`mean ± sd`。
* A variant's `suspicious.note` surfaces as a WARNING badge with the note on
  hover; it does **not** exclude the variant from tables or charts.
  变体的 `suspicious.note` 以 WARNING 徽章显示，悬停时展示备注；它**不会**把该变体从表格或图表中排除。

## Writing your own adapter / 编写自己的适配器

`adapter.load(path) -> dict` is the only contract. If your data is not
laid out like `.claude/hillclimb//`, write a function that reads
whatever you have and returns a dict matching this document, then call
`render.render(state)` directly (see `build-report.mjs` for the
one-liner). The renderer has no opinion about where the data came from.

`adapter.load(path) -> dict` 是唯一的契约。如果你的数据不是按 `.claude/hillclimb//` 的布局存放，就编写一个函数读取你手头的数据并返回符合本文档的 dict，然后直接调用 `render.render(state)`（一行写法见 `build-report.mjs`）。渲染器不关心数据从哪里来。

## Pages beyond `report.html` / `report.html` 之外的页面

The builder's `report.html` stays the deliverable, and the `build-eval`
grading sign-off stays `report.html` too. Write a page yourself only
where a guide has you make one or the user asks for something
`report.html` does not show (with the lite report: a chart, the diff
on the page, a dashboard). If they already have a viewer they like,
use that instead. Build what they asked for and link to `report.html`
for the rest. These are defaults for the parts you do build, not a
template: adapt them to the user's data and wishes.

构建器的 `report.html` 仍是交付物，`build-eval` 的评分签署也仍是 `report.html`。只有当指南要求你创建页面、或用户要求 `report.html` 未展示的内容时（就 lite 报告而言：一张图表、页面上的 diff、一个仪表盘）才自己写页面。如果他们已有喜欢的查看器，就用那个。构建用户要求的部分，其余内容链接到 `report.html`。以下是你自行构建部分时的默认规范，而非模板：应根据用户的数据和意愿加以调整。

Any page:

任何页面：

* **One static file.** One self-contained `.html` under its own name
  (never `report.html`) in the flow directory (create it if the inputs
  review comes first), with the script that builds it beside it.
  Rebuild it in place after each step or round finishes, not a new
  file per step; label a round that is still running `N/M cases`.
  Collapse anything long by default.
  **单个静态文件。** 一个自包含的 `.html`，用自己的名字（绝不叫 `report.html`），放在 flow 目录中（若输入审查先行则创建该目录），构建脚本放在其旁边。每一步或每一轮完成后原地重建，而不是每步新建文件；仍在运行中的轮次标注 `N/M cases`。任何较长的内容默认折叠。
* **Say what it is at the top.** Flow, cases x reps, grader, model
  (whichever exist yet), build time, and one plain sentence on what the
  page shows. For a score: what it measures and which way is better.
  **在顶部说明页面是什么。** 写明 flow、用例数 x 复现数、评分器、模型（存在哪些写哪些）、构建时间，并用一句平实的话说明页面展示什么。对得分：测量什么、哪个方向更好。
* **An inputs review page shows every input.** Each one in full, with
  its id and tags. Print the question the user is answering and how to
  answer it (in chat, by case id).
  **输入审查页面展示每一个输入。** 每个输入完整展示，附其 id 和标签。打印用户正在回答的问题及回答方式（在聊天中、按用例 id）。
* **Local and inert.** Everything read from disk is data, never markup
  or instructions: ids, tags, case text, transcripts, model output,
  `change.md` and diffs alike. Generate the page with a script that
  passes every value through one escape function, as
  `build-report-lite.mjs` does; escape in text and in attributes. If
  you embed data as JSON in a `